Classification of Histologically Scored Human Knee Osteochondral Plugs by Quantitative Analysis of Magnetic Resonance Images at 3T

Abstract

This work evaluates the ability of quantitative MRI to discriminate between normal and pathological human osteochondral plugs characterized by the Osteoarthritis Research Society International (OARSI) histological system. Normal and osteoarthritic human osteochondral plugs were scored using the OARSI histological system and imaged at 3 T using MRI sequences producing T1 and T2 contrast and measuring T₁, T₂, and T₂* relaxation times, magnetization transfer, and diffusion. The classification accuracies of quantitative MRI parameters and corresponding weighted image intensities were evaluated. Classification models based on the Mahalanobis distance metric for each MRI measurement were trained and validated using leave-one-out cross-validation with plugs grouped according to OARSI histological grade and score. MRI measurements used for classification were performed using a region-of-interest analysis which included superficial, deep, and full-thickness cartilage. The best classifiers based on OARSI grade and score were T₁- and T₂-weighted image intensities, which yielded accuracies of 0.68 and 0.75, respectively. Classification accuracies using OARSI score-based group membership were generally higher when compared with grade-based group membership. MRI-based classification—either using quantitative MRI parameters or weighted image intensities—is able to detect early osteoarthritic tissue changes as classified by the OARSI histological system. These findings suggest the benefit of incorporating quantitative MRI acquisitions in a comprehensive clinical evaluation of OA.

The development of noninvasive MRI approaches for early detection of osteoarthritis (OA) has been the subject of intense activity. Quantitative MRI parameters (e.g., T₂ and T_1ρ relaxation times, magnetization transfer, apparent diffusion coefficient, and displacement encoding) have been shown to be sensitive to cartilage extracellular matrix constituents.^1–5 Additionally, advances in MRI acquisition have improved anatomical coverage and image contrast for evaluation of cartilage volume, cartilage thickness, and surface defects.^6,7 These advances seek to provide faster acquisition of images with contrasts such as T₁-, T₂-, and proton density-weighting.⁸ Despite these developments early detection of OA remains an elusive goal, in part because the only imaging definition of OA is based on radiography.⁹ Furthermore, there remains great debate over the definition of OA between clinicians and basic scientists.¹⁰ For an MRI-based detection of early OA, such a definition is needed to determine which imaging sequences and measurements provide the best prognostic and diagnostic accuracy.

A great deal of work has established relationships between quantitative MRI parameters and extracellular matrix properties. For instance, to examine the sensitivity of T₂ and T_1ρ to proteoglycan and collagen content, comparisons have been made with quantitative biochemical measurements in model articular cartilage systems¹¹ and in both normal and osteoarthritic articular cartilage.^5,11 Immunohistochemical staining has also demonstrated qualitative correspondence between quantitative MRI parameter values and matrix composition in normal and degraded human articular cartilage.¹¹ T₂ anisotropy measurements have been shown to correspond with articular cartilage collagen orientation as measured with polarized light microscopy (PLM) across age and species including in human articular cartilage.¹² T₂* measurements in human articular cartilage have been shown to correlate with both histopathology, based on Mankin grades, and PLM-based birefringence grades, suggesting sensitivity to changes in matrix composition and structure occurring during osteoarthritis.¹³

These studies represent the foundation of work establishing MRI markers for detection of early cartilage changes that likely prognosticate radiographic OA. Many of these parameters show measurable differences between normal and diseased tissue. Unfortunately, most of these studies only report group mean differences rather than diagnostic accuracy (e.g., true positive and true negative rates), limiting any conclusions as to the diagnostic value of these measurements.¹⁴ Recently, we examined the classification accuracy of several quantitative MRI parameters in enzymatically degraded bovine nasal cartilage (BNC) —a model system for articular cartilage.^15,16 This work detailed a variety of approaches for building classification models and showed differences in the accuracy of several quantitative MRI parameters in classifying cartilage to either control or enzymatically-degraded groups. T₁ proved to be the most accurate MRI parameter at 9.4 T for discriminating between normal and severely degraded BNC, with sensitivity and specificity of 0.90 and 0.94, respectively.¹⁵ In a recent study on the classification of normal and enzymatically degraded BNC at 3 T, T₁ showed comparable classification accuracy.¹⁶ These previous studies utilized the Euclidian distance metric of classification, which assigns group membership of a sample based on the proximity of its parameter value to the group means. Other preliminary work performed using ultra-high field preclinical MRI has demonstrated the potential for MRI parameters to discriminate between various stages of OA using receiver operating characteristic (ROC) analysis.¹⁷

Here, we extend our previous classification work to human osteochondral plugs to examine the diagnostic accuracy of several quantitative MRI measurements at clinical field strength and physiological body temperature.^15,16 Plugs were scored for OA severity using the Osteoarthritis Research Society International (OARSI) histological system, a gold standard for tissue characterization.¹⁸ Classification models were constructed using the Mahalanobis distance metric, a variant of the Euclidean distance metric, which normalizes the distance between the sample parameter value and the group mean parameter value by the standard deviation of the group. This normalization allows the model to account for potential differences between group parameter distributions, as would be expected between healthy and diseased cartilage.⁹ We restricted our MRI acquisition methods to those which (1) previously demonstrated sensitivity to cartilage matrix changes and (2) do not necessitate the use of exogenous contrast agents, to maintain broad applicability of our results. All MRI acquisition methods are widely-available on clinical scanners with the exception of T₁, which was included due to its high classification performance exhibited in BNC.¹⁵ In addition to measurement of quantitative MRI tissue parameters and in recognition of the wide-spread use of weighted images in clinical practice and research studies such as the Osteoarthritis Initiative (OAI),¹⁹ we also analyzed specific weighted images.

MATERIALS AND METHODS

Tissue Acquisition

Tissue from patients (9 males and 10 females; average age = 66 years; age range = 44–79 years) undergoing elective knee joint replacement were obtained according to a protocol approved by the Institutional Review Board. Additionally, knees from the donated body program (2 females; ages 52 and 56 years) with no visible signs of OA based on radiographic screening were obtained. Patient information is summarized in Supplementary Material 1. Osteochondral plugs with 5 mm diameter and 5 mm height were harvested from defined femur locations (Fig. 1; LA = lateral anterior (n = 24), LP = lateral posterior (n = 26), MA = medial anterior (n = 10), and MP= medial posterior (n = 20)). Note that LA and MA locations were within relatively high contact pressure joint regions, whereas locations LP and MP were within low contact pressure joint regions.^20,21 Within a given subject joint, samples showed a continuum of minimum to advanced damage, evidenced by gradients of mechanical (e.g., surface friction) and structural measures over the femoral locations assessed in additional, site-matched samples.²¹ In some cases, samples from multiple compartments of a given subject were available; these were considered independent in our analysis as histology was used as the gold standard of structural and compositional disease progression. Samples dedicated for histology (n = 40) were fixed immediately after harvest to limit degradation post-surgery. Adjacent paired samples from each location exhibiting comparable gross pathology were harvested for MRI assessment (n = 40), numerically labeled, flash-frozen, and stored at −80°C until imaged.

Figure 1.

Standardized femur harvest locations (LA = lateral anterior, LP = lateral posterior, MA = medial anterior, and MP = medial posterior). From each location, cylindrical osteochondral samples (diameter = 5 mm; height = 5 mm) were harvested for histology (1) and MRI (2).

Histopathology

The 40 histology-dedicated osteochondral plugs were fixed, decalcified, and embedded, and sections were stained by hematoxylin and eosin, and toluidine blue. The OARSI histopathology assessment was performed by two independent, expert observers.¹⁸ The OARSI score (0–24) is an overall assessment of the severity and extent of OA and is calculated as the product of the grade and stage. A grade (0–6) indicates the biologic depth progression of OA in the cartilage. The stage (0–4) reflects the percent of cartilage surface area affected by OA.

MRI Sample Preparation

The 40 imaging-dedicated osteochondral plugs were imaged in a susceptibility-matched polyetherimide (ULTEM) sample holder containing four vertical sample wells 9 mm in diameter and 8.5cm deep. Three wells contained samples submerged in Fluorinert^® FC-77 (Sigma–Aldrich, St. Louis, MO). The fourth well was filled with Dulbecco’s phosphate-buffered saline (DPBS; Invitrogen), a signal intensity (SI) standard. During handling and transport, all samples were maintained at 4°C to minimize sample degradation prior to imaging.

MRI Acquisition

Imaging data were acquired using a 3 T Philips Achieva MRI scanner (Philips Healthcare, Best, The Netherlands) equipped with an eight channel SENSE knee coil. The sample holder was positioned perpendicular to the B₀ static magnetic field; this oriented the cartilage surfaces to be near-parallel to B₀ (Fig. 2). Sagittal imaging slices were placed to ensure a central bisecting slice through each well of samples, therefore requiring two slices to image samples from all four wells using 2D sequences.

Figure 2.

Proton density weighted images (2D spin echo with TE = 10 ms) of the ULTEM sample holder containing six human articular cartilage plugs per well and a DPBS standard. These images are representative of the slice orientation acquired for each image acquisition; a)slice through well 1 and 2; b)slice through well 3 and 4. The sample orientation is designated relative to head-first supine patient positioning with “anterior” and “left” indicated on the images with the letters “A” and “L,” respectively, and the static magnetic field B₀ oriented into the image plane. Cartilage (arrow) was delineated from subchondral bone.

One hour before imaging, samples were warmed to 37 ± 0.1°C in the magnet using an MRI-compatible heating module (SA Instruments, Stony Brook, NY). Sample temperature was monitored using an MRI-compatible thermometer attached to the sample holder and maintained at 37 ± 0.1°C for the duration of image acquisition (approximately 2.5 h).

The following MRI acquisitions were used to yield quantitative parameter or weighted image intensity MRI measurements, including those used in the OAI.¹⁹ The acquisition parameters were chosen based on the protocols available on our clinical 3 T MR system and were then optimized to obtain adequate signal-to-noise ratio (SNR) and spatial resolution for quantitative MRI tissue parameter calculation for the ex vivo specimens.

Images for calculation of T₁ relaxation times were obtained using a 2D Look-Locker sequence with echo-planar imaging (EPI) readout with EPI factor = 3, TR/TE = 6 s/5 ms, 55 inversion times (TI) ranging from 18ms to 2757 ms, and flip angle (FA) = 14°. Two 4 mm thick sagittal slices were acquired using bandwidth (BW) = 17.5 kHz, field-of-view (FOV) = 75 × 44.5 mm² (vertical × horizontal), acquisition matrix (MTX) = 120 × 66, voxel size (VOX) = 0.63 × 0.67 mm², and number of signals averaged (NSA) = 2, for a total scan time of 7.7 min. T₁-weighted (T₁W) images were obtained using a 3D fast low angle shot (FLASH) sequence with TR/TE = 20 ms/7.57ms, FA = 13°, partial Fourier factor = 0.75, BW = 31.1 kHz, FOV = 75 × 45 × 24 mm³, MTX = 240 × 144 × 16, VOX = 0.31 × 0.31 × 1.50 mm³, and NSA = 1 for a scan time = 14 s. Though this is not the conventional clinical spin echo or turbo spin echo T₁W acquisition, it is often used for cartilage segmentation and comparable to that used in the OAI.

Images for calculation of T₂ relaxation times were obtained using two methods. First, a 3D multi-echo spin echo sequence with TR/TE = 767 ms/12ms and an echo train length (ETL) = 30 was applied with BW = 28.2 kHz, FOV = 75 × 45 × 23 mm³, MTX = 188 × 78 × 7, VOX = 0.40 × 0.58 × 3.30 mm³, and NSA = 1 for a scan time of 13.0 min. Second, a conventional 2D multi-slice multi-echo spin echo sequence comparable to the OAI protocol was applied with TR/TE = 2700 ms/10 ms, ETL = 7, BW = 59.8 kHz, FOV = 75 × 45 mm², MTX = 240 × 101, VOX = 0.31 × 0.45 mm², and NSA = 1 for a scan time of 4.5 min.

Images for calculation of T₂* relaxation times were obtained using a 2D gradient echo sequence to acquire two 3.5 mm thick slices with TE₁ = 1.5 ms, ΔTE = 4.2 ms, TR = 2s, FA = 25°, ETL = 30 BW = 98.9 kHz, FOV = 75 × 45 mm², MTX = 152 × 73, VOX = 0.49 × 0.62 mm², and NSA = 2 for a scan time of 4.9 min.

Images for apparent diffusion coefficient (ADC) measurements were obtained with a 2D diffusion prepared gradient echo with EPI readout to acquire two 4 mm thick slices with TR/TE = 2 s/62 ms, EPI factor = 3, Δ = 25.3 ms, δ = 12.4ms, BW = 12kHz, FOV = 75 × 43.75 mm², MTX = 96 × 43, VOX = 0.78 × 0.95 mm², NSA = 1, and b values of 0, 333, 666, 1000, 1333, 1666 and 2000 s/mm2 applied in three orthogonal diffusion-sensitizing gradient directions for a scan time of 15.5 min.

Magnetization transfer (MT) weighted images were obtained with a 2D gradient echo sequence with TR/TE = 517 ms/2.4 ms and FA = 25°, preceded by a series of 50 ms duration, 1 kHz off-resonance saturation block pulses with B₁ = 2.15 µT to acquire two 5 mm slices. The number of saturation pulses was varied to give saturation times (T_sat) ranging from 50 to 200 ms. Other acquisition parameters included BW = 98.9 kHz, FOV = 75 × 45 mm², MTX = 152 × 73, VOX = 0.49 × 0.62 mm², and NSA = 2 for a scan time of 10 min. Fat shift was posterior in all sequences (towards bone) and the vertical voxel dimension was through the cartilage depth.

MRI Data Analysis

Three regions-of-interest (ROI) were selected in each osteochondral plug using a reference proton density image to allow for analysis of depth-wise variations of MRI signal as shown in Figure 3. The first ROI was drawn selecting the top one-third of cartilage tissue depth from the surface. Although we will refer to this as the “superficial” ROI, it is not to be confused with the histologically-identified superficial zone. The second ROI was drawn selecting the remaining two-thirds of the cartilage tissue depth and will be referred to as the “deep” ROI—again, not to be confused with the histologically-identified deep zone. The third ROI was constructed using the pixels from both the superficial and deep ROIs, and is referred to as the full-thickness ROI. Each ROI was drawn to exclude the pixels at the cartilage margins which might be contaminated with either subchondral bone or the surrounding Fluorinert bath. It is worth noting that non-protonated Fluorinert fluid does not produce any MR signal in these experiments and thus would not affect calculated MR parameter values through partial volume averaging. Each ROI contained a minimum of 3 pixels.

Figure 3.

Example of superficial (top) and deep (bottom) ROI selections drawn on a proton density weighted image (2D spin echo with TE = 10 ms) to exclude pixels at the cartilage margins. The full-thickness ROI was defined as the average SI from the superficial and deep pixels combined.

Quantitative MRI parameters T₁, T₂, T₂*, and ADC were obtained from fits of the average ROI SI from each dynamic scan to a three-parameter monoexponential model including baseline offset.^22,23 For T₂ analyses of 3D spin echo data, all 30 echo images from the echo train were used, however only images with echo times from 20 to 70 ms were fit for 2D spin echo data with the 10ms echo image excluded due to pulse imperfections.²⁴ Magnetization transfer ratio (MTR) was calculated as 1 − M_sat/M₀ where M_sat is the SI for a T_sat of 200ms and M₀ is the SI in the absence of off-resonance saturation. For diffusion-weighted (DW), MT-weighted (MTW), T₁W, and T₂-weighted (T₂W) SI measurements, which were also individually used for classification analysis, the average cartilage ROI SI was normalized to the average SI from the DPBS solution in the same weighted image for standardization between imaging sessions.

Preliminary repeatability experiments were used to assess inter-session variability for all MRI sequences. These resulted in coefficients of variation of <7% for relaxation times and weights, <14% for ADC measurements and weights, and <2% for MT measurements.

Univariate Classification Analysis

Two separate analyses were performed using either OARSI grade or score values for assignment of sample membership to “normal” or “diseased” groups.

Grade-based analysis defined samples with values >2 to the diseased group, resulting in 23 normal samples and 17 diseased samples. Score-based analysis defined samples with values >6 to the diseased group, resulting in 20 normal samples and 20 diseased samples. These definitions of normal tissue include grades and scores consistent with the cartilage obtained from the donated body tissue, for which diagnostic signs of OA were exclusion criteria.

For each MRI measurement, a classification model was constructed using the Mahalanobis distance metric. Using the parameter value of the i^th sample (p_i), the Mahalanobis distance metric classifies that sample (S_i) to the normal (Norm) or diseased (Dis) group based on distance from their means (μ) normalized by their standard deviations (σ) as shown in the following assignment rules:

$$|\frac{p_i-{\mu }_{\mathit{\text{Norm}},i}}{{\sigma }_{\mathit{\text{Norm}},i}}|<|\frac{p_i-{\mu }_{\mathit{\text{Dis}},i}}{{\sigma }_{\mathit{\text{Dis}},i}}|\Rightarrow S_i\in \mathit{\text{Norm}}$$ (1)

$$|\frac{p_{i-{\mu }_{\mathit{\text{Dis}},i}}}{{\sigma }_{\mathit{\text{Dis}},i}}|<|\frac{p_{i-{\mu }_{\mathit{\text{Norm}},i}}}{{\sigma }_{\mathit{\text{Norm}},i}}|\Rightarrow S_i\in \mathit{\text{Dis}}$$ (2)

The classification analysis was performed using an in-house Matlab script (The MathWorks, Inc. Natick, MA) utilizing a leave-one-out cross-validation scheme for model training and validation.²⁵ Classification results are reported in terms of sensitivity (rate of true positives), specificity (rate of true negatives), and accuracy (average of sensitivity and specificity). In the case of grade-based assignment with unequal group sizes, 100 random selections of 17 out of the 23 normal samples were made to equalize group sizes. Leave-one-out cross-validation was performed 100 times and classification results are reported as the mean ± standard deviation over all trials. Score-based assignment produced groups of equal size, permitting a single leave-one-out cross-validation procedure.

Statistical Analysis

Data are reported as mean ± standard deviation. Data for normal and diseased groups were compared using a two-tailed t-test assuming unequal variance. Statistically significant differences are denoted as p<0.1 (*), p<0.05 (**), or p<0.01 (***).

RESULTS

For grade-based analysis the average grade, stage, and score values were 1.6 ± 0.4, 2.4 ± 1.0, and 4.1 ± 2.3 in the normal group and 2.9 ± 0.8, 3.4 ± 0.6, and 9.9 ± 3.8 in the diseased group. For the grade-based analysis, no diseased sample exceeded a grade of 4.5, representing denudation of the cartilage surface. For score-based analysis the average grade, stage, and score values were 1.6 ± 0.4, 2.1 ± 0.9, and 3.6 ± 2.0 for the normal group and 2.7 ± 0.8, 3.5 ± 0.5, and 9.5 ± 3.6 in the diseased group. For score-based analysis, no diseased sample exceeded a score of 18, which would be comparable to a grade of 4.5 and stage of 4. Complete grade, stage, and score information by sample is included in Supplementary Materials 1.

Tables 1 and 2 show the means and standard deviations for all MRI measurements in the superficial, deep, and full-thickness ROIs for samples assigned to normal and diseased groups as determined by OARSI grade and score, respectively. The majority of these did not show statistically significant differences between groups. This was especially true when samples were classified according to grade rather than score. When samples were grouped by grade, T₂, MTR, T₁W contrast, and 2D spin echo contrasts with TE > 10 ms showed differences with varying degrees of statistical significance, primarily in the full-thickness and superficial ROIs. ADC and DW (b = 1333 and 2000 s/mm²) showed significant differences in the superficial ROIs. More statistically significant differences were observed between groups defined by OARSI score, with these differences observed for T₁, T₂, MTW, T₁W contrast, and all 2D spin echo contrasts.

Table 1.

MRI Parameters of Human Osteochondral Plugs by OARSI Grade

		ROI
MRI Measurement	Group Assignment	Superficial	Deep	Full
T1 (ms)	Normal	855.2 ± 70.1	817.5 ± 75.7	834.6 ± 69.3
	Diseased	875.3 ± 70.1	827.4 ± 77.8	852.5 ± 65.4
T2 3D (ms)	Normal	42.7 ± 6.1	37.2 ± 8.3	39.7 ± 6.8
	Diseased	47.4 ± 8.9*	40.8 ± 5.1	43.9 ± 6.1**
T2* (ms)	Normal	15 ± 7	16 ± 4	16± 4
	Diseased	13 ± 6	15 ± 5	14 ± 5
ADC (×10−3 mm2/s)	Normal	2.70 ± 0.43	2.61 ± 0.44	2.66 ± 0.41
	Diseased	2.47 ± 0.36*	2.55 ± 0.44	2.49 ± 0.34
MTR	Normal	0.46 ± 0.07	0.48 ± 0.07	0.47 ± 0.07
	Diseased	0.40 ± 0.10**	0.43 ± 0.09*	0.42 ± 0.09**
T1W	Normal	1.21 ± 0.16	1.19 ± 0.15	1.19 ± 0.15
	Diseased	1.09 ± 0.18*	1.11 ± 0.14	1.10 ± 0.16*
T2 2D (ms)	Normal	33.4 ± 5.7	31.9 ± 6.6	32.4 ± 4.7
	Diseased	35.6 ± 5.1	31.0 ± 4.8	33.4 ± 4.1
T2W (TE = 10 ms)	Normal	1.11 ± 0.11	1.09 ± 0.09	1.10 ± 0.09
	Diseased	1.16 ± 0.16	1.13 ± 0.15	1.15 ± 0.15
T2W (TE = 20 ms)	Normal	0.95 ± 0.10	0.89 ± 0.09	0.92 ± 0.08
	Diseased	0.99 ± 0.14	0.94 ± 0.13	0.96 ± 0.13*
T2W (TE = 30 ms)	Normal	0.76 ± 0.09	0.69 ± 0.10	0.72 ± 0.09
	Diseased	0.81 ± 0.13	0.74 ± 0.12	0.77 ± 0.12*
T2W (TE = 40 ms)	Normal	0.64 ± 0.08	0.57 ± 0.10	0.60 ± 0.08
	Diseased	0.70 ± 0.12*	0.62 ± 0.12	0.66 ± 0.12*
T2W (TE = 50 ms)	Normal	0.52 ± 0.08	0.46 ± 0.10	0.49 ± 0.08
	Diseased	0.59 ± 0.12*	0.51 ± 0.11	0.54 ± 0.10*
T2W (TE = 60 ms)	Normal	0.46 ± 0.08	0.39 ± 0.10	0.42 ± 0.08
	Diseased	0.52 ± 0.11*	0.46 ± 0.10	0.48 ± 0.10*
T2W (TE = 70 ms)	Normal	0.38 ± 0.07	0.32 ± 0.09	0.34 ± 0.07
	Diseased	0.44 ± 0.11*	0.37 ± 0.09*	0.40 ± 0.09*
DW (b = 333s/mm2)	Normal	0.49 ± 0.13	0.54 ± 0.16	0.52 ± 0.13
	Diseased	0.52 ± 0.11	0.51 ± 0.14	0.52 ± 0.12
DW (b = 666s/mm2)	Normal	0.75 ± 0.23	0.82 ± 0.28	0.80 ± 0.23
	Diseased	0.84 ± 0.24	0.90 ± 0.29	0.88 ± 0.26
DW (b = 1000 s/mm2)	Normal	1.22 ± 0.43	1.45 ± 0.63	1.32 ± 0.48
	Diseased	1.30 ± 0.53	1.39 ± 0.48	1.37 ± 0.45
DW (b = 1333 s/mm2)	Normal	1.05 ± 0.39	1.15 ± 0.57	1.13 ± 0.44
	Diseased	1.25 ± 0.31*	1.38 ± 0.36	1.31 ± 0.31
DW (b = 1666 s/mm2)	Normal	1.15 ± 0.29	1.07 ± 0.43	1.13 ± 0.29
	Diseased	1.02 ± 0.25	1.11 ± 0.36	1.06 ± 0.24
DW (b = 2000 s/mm2)	Normal	1.05 ± 0.28	1.16 ± 0.45	1.08 ± 0.32
	Diseased	0.89 ± 0.21**	1.07 ± 0.27	0.99 ± 0.21
MTW (Tsat = 50 ms)	Normal	0.55 ± 0.07	0.55 ± 0.08	0.55 ± 0.07
	Diseased	0.53 ± 0.09	0.53 ± 0.08	0.53 ± 0.08
MTW (Tsat = 100 ms)	Normal	0.56 ± 0.05	0.55 ± 0.06	0.56 ± 0.05
	Diseased	0.57 ± 0.05	0.57 ± 0.07	0.57 ± 0.06
MTW (Tsat = 150 ms)	Normal	0.49 ± 0.05	0.47 ± 0.05	0.48 ± 0.04
	Diseased	0.49 ± 0.05	0.49 ± 0.05	0.49 ± 0.05
MTW (Tsat = 200 ms)	Normal	0.43 ± 0.04	0.41 ± 0.05	0.42 ± 0.04
	Diseased	0.44 ± 0.04	0.43 ± 0.05	0.44 ± 0.05

MRI parameters (mean ± standard deviation) of human osteochondral plugs grouped by OARSI Grade: Normal (< = 2), n = 23; Diseased (>2), n = 17.

^*p < 0.1 normal vs. diseased and

^**p < 0.05 normal vs. diseased.

Table 2.

MRI Parameters of Human Osteochondral Plugs by OARSI Score

		ROI
MRI Measurement	Group Assignment	Superficial	Deep	Full
T1 (ms)	Normal	844.8 ± 68.3	801.6 ± 74.9	821.5 ± 68.3
	Diseased	882.7 ± 58.9*	841.9 ± 73.0*	863.0 ± 61.3*
T2 3D (ms)	Normal	41.7 ± 5.0	36.4 ± 7.9	38.8 ± 6.0
	Diseased	47.7 ± 8.8**	41.2 ± 5.9**	44.1 ± 6.6**
T2* (ms)	Normal	15 ± 7	16 ± 4	16 ± 4
	Diseased	13 ± 6	15 ± 5	14 ± 5
ADC (×10−3 mm2/s)	Normal	2.71 ± 0.46	2.64 ± 0.45	2.68 ± 0.43
	Diseased	2.50 ± 0.33	2.52 ± 0.42	2.50 ± 0.33
MTR	Normal	0.46 ± 0.07	0.48 ± 0.08	0.47 ± 0.07
	Diseased	0.42 ± 0.10	0.44 ± 0.08	0.43 ± 0.09
T1W	Normal	1.21 ± 0.17	1.18 ± 0.16	1.19 ± 0.16
	Diseased	1.11 ± 0.17*	1.13 ± 0.14	1.12 ± 0.15
T2 2D (ms)	Normal	32.6 ± 3.9	30.8 ± 6.8	31.5 ± 4.3
	Diseased	35.3 ± 5.4**	31.8 ± 4.4	33.8 ± 3.7**
T2W (TE = 10 ms)	Normal	1.08 ± 0.10	1.06 ± 0.10	1.07 ± 0.09
	Diseased	1.18 ± 0.15**	1.16 ± 0.13**	1.17 ± 0.13**
T2W (TE = 20 ms)	Normal	0.92 ± 0.09	0.86 ± 0.09	0.89 ± 0.08
	Diseased	1.01 ± 0.12**	0.96 ± 0.11***	0.98 ± 0.11***
T2W (TE = 30 ms)	Normal	0.73 ± 0.09	0.66 ± 0.10	0.69 ± 0.08
	Diseased	0.83 ± 0.12***	0.76 ± 0.10***	0.79 ± 0.10***
T2W (TE = 40 ms)	Normal	0.62 ± 0.08	0.55 ± 0.10	0.58 ± 0.08
	Diseased	0.72 ± 0.11***	0.64 ± 0.11***	0.68 ± 0.10***
T2W(TE = 50 ms)	Normal	0.50 ± 0.07	0.43 ± 0.09	0.47 ± 0.07
	Diseased	0.59 ± 0.10***	0.53 ± 0.10***	0.56 ± 0.09***
T2W (TE = 60 ms)	Normal	0.44 ± 0.07	0.37 ± 0.09	0.40 ± 0.07
	Diseased	0.52 ± 0.10***	0.46 ± 0.10***	0.49 ± 0.09***
T2W (TE = 70 ms)	Normal	0.36 ± 0.07	0.29 ± 0.08	0.33 ± 0.06
	Diseased	0.44 ± 0.10***	0.38 ± 0.09***	0.41 ± 0.09***
DW (b = 333 s/mm2)	Normal	0.47 ± 0.11	0.50 ± 0.13	0.49 ± 0.11
	Diseased	0.54 ± 0.12*	0.55 ± 0.17	0.55 ± 0.14
DW (b = 666 s/mm2)	Normal	0.73 ± 0.21	0.78 ± 0.23	0.77 ± 0.19
	Diseased	0.85 ± 0.25*	0.93 ± 0.32	0.89 ± 0.27
DW (b = 1000 s/mm2)	Normal	1.20 ± 0.35	1.40 ± 0.49	1.30 ± 0.37
	Diseased	1.31 ± 0.57	1.45 ± 0.64	1.38 ± 0.55
DW (b = 1333 s/mm2)	Normal	1.04 ± 0.32	1.16 ± 0.37	1.14 ± 0.31
	Diseased	1.23 ± 0.39	1.33 ± 0.60	1.28 ± 0.46
DW (b = 1666 s/mm2)	Normal	1.09 ± 0.29	1.06 ± 0.33	1.10 ± 0.20
	Diseased	1.11 ± 0.28	1.11 ± 0.46	1.11 ± 0.33
DW (b = 2000 s/mm2)	Normal	1.02 ± 0.25	1.11 ± 0.38	1.04 ± 0.26
	Diseased	0.94 ± 0.28	1.13 ± 0.39	1.04 ± 0.30
MTW (Tsat = 50 ms)	Normal	0.54 ± 0.06	0.53 ± 0.08	0.54 ± 0.07
	Diseased	0.54 ± 0.09	0.55 ± 0.08	0.55 ± 0.08
MTW (Tsat = 100 ms)	Normal	0.55 ± 0.05	0.54 ± 0.06	0.55 ± 0.05
	Diseased	0.58 ± 0.05	0.58 ± 0.06*	0.58 ± 0.06
MTW (Tsat = 150 ms)	Normal	0.48 ± 0.04	0.46 ± 0.05	0.47 ± 0.04
	Diseased	0.50 ± 0.05	0.49 ± 0.05**	0.50 ± 0.05*
MTW (Tsat = 200 ms)	Normal	0.42 ± 0.04	0.41 ± 0.05	0.42 ± 0.04
	Diseased	0.44 ± 0.04	0.44 ± 0.05*	0.44 ± 0.04*

MRI parameters (mean ± standard deviation) of human osteochondral plugs grouped by OARSI Score: Normal (< = 6), n = 20; Diseased (>6), n = 20.

^*p < 0.1 normal vs. diseased;

^**p < 0.05 normal vs. diseased;

^***p < 0.01 normal vs. diseased.

Table 3 shows the sensitivity, specificity, and accuracy for the top three MRI classifiers for each ROI for grade-based analysis. Training classification results are shown in addition to validation classification results, with training results reflecting the maximum ability for a given parameter to classify. Full-thickness T₁W contrast showed the highest grade-based classification accuracy with validation sensitivity of 0.73 ± 0.04, specificity of 0.64 ± 0.06, and accuracy of 0.69 ± 0.05. T₁W contrast in the deep zone showed similar accuracy with sensitivity of 0.68 ± 0.04, specificity of 0.67 ± 0.05, and accuracy of 0.67 ± 0.03. While the accuracies of full-thickness and deep zone analysis were comparable, the full-thickness measurements showed a higher sensitivity and lower specificity than the deep zone measurements. DW (b = 2000 s/mm²) measurements in the superficial ROI showed a similar accuracy to that observed for deep and full-thickness ROIs using T₁W contrast though their specificity was only 0.59 ± 0.04.

Table 3.

Univariate Classification of Human Osteochondral Plugs by OARSI Grade.

		Training Set			Validation Set
ROI	MRI Measurement	Sensitivity	Specificity	Accuracy	Sensitivity	Specificity	Accuracy
Superficial	DW (b = 2000 s/mm2)	0.76 ± 0.00	0.60 ± 0.04	0.68 ± 0.02	0.75 ± 0.02	0.59 ± 0.04	0.67 ± 0.03
	MTR	0.58 ± 0.01	0.75 ± 0.04	0.67 ± 0.02	0.58 ± 0.02	0.73 ± 0.05	0.66 ± 0.03
	T2 3D (ms)	0.62 ± 0.07	0.71 ± 0.04	0.67 ± 0.04	0.58 ± 0.07	0.69 ± 0.06	0.63 ± 0.04
Deep	T1W	0.70 ± 0.02	0.67 ± 0.04	0.68 ± 0.02	0.68 ± 0.04	0.67 ± 0.05	0.67 ± 0.03
	MTR	0.52 ± 0.06	0.76 ± 0.04	0.64 ± 0.04	0.51 ± 0.06	0.76 ± 0.04	0.64 ± 0.04
	T2W (TE = 70 ms)	0.61 ± 0.03	0.65 ± 0.06	0.63 ± 0.05	0.56 ± 0.06	0.65 ± 0.06	0.61 ± 0.05
Full	T1W	0.75 ± 0.02	0.66 ± 0.05	0.71 ± 0.03	0.73 ± 0.04	0.64 ± 0.06	0.69 ± 0.05
	T2 3D (ms)	0.58 ± 0.02	0.77 ± 0.04	0.68 ± 0.02	0.53 ± 0.02	0.77 ± 0.04	0.65 ± 0.02
	MTR	0.53 ± 0.01	0.78 ± 0.04	0.65 ± 0.03	0.53 ± 0.02	0.78 ± 0.04	0.65 ± 0.03

Univariate classification of human osteochondral plugs grouped by OARSI grade for the top three performing MRI measurements in superficial, deep, and full-thickness ROIs. Sensitivity, specificity, and accuracy are reported as the means and standard deviations over 100 random selections of training sets and validation sets of equal size normal (n = 23) and diseased (n = 17) groups using the Mahalanobis distance metric and leave-one-out cross-validation method.

Table 4 shows the sensitivity, specificity, and accuracy for the top three MRI classifiers in each ROI for score-based analysis. Full-thickness measurements of T₂W contrast with TE = 60 ms showed the highest classification accuracy with validation sensitivity of 0.75, specificity of 0.75, and accuracy of 0.75. 3D spin echo T₂ and 2D T₂W contrasts comprised the top three classifiers for superficial, deep, and full-thickness ROIs with no accuracy falling below 0.70.

Table 4.

Univariate Classification of Human Osteochondral Plugs by OARSI Score

		Training Set			Validation Set
ROI	MRI Measurement	Sensitivity	Specificity	Accuracy	Sensitivity	Specificity	Accuracy
Superficial	T2 (ms)	0.73	0.73	0.73	0.70	0.75	0.73
	T2W (TE = 70 ms)	0.76	0.70	0.73	0.75	0.70	0.73
	T2W (TE = 50 ms)	0.70	0.70	0.70	0.70	0.70	0.70
Deep	T2W (TE = 60 ms)	0.70	0.75	0.73	0.70	0.75	0.73
	T2W (TE = 70 ms)	0.69	0.75	0.72	0.70	0.75	0.73
	T2W (TE = 10 ms)	0.65	0.79	0.72	0.65	0.80	0.73
Full	T2W (TE = 60 ms)	0.74	0.77	0.76	0.75	0.75	0.75
	T2W (TE = 70 ms)	0.73	0.75	0.74	0.70	0.75	0.73
	T2W (TE = 10 ms)	0.73	0.75	0.74	0.70	0.75	0.73

Univariate classification of human osteochondral plugs grouped by OARSI score for the top three performing MRI measurements in superficial, deep, and full-thickness ROIs. Sensitivity and specificity are reported for the training set and validation set drawn from normal (n = 20) and diseased (n = 20) groups using the Mahalanobis distance metric and leave-one-out cross-validation method.

Overall, the highest classification accuracies were obtained in full-thickness ROIs in samples grouped according to OARSI score. Only in a few instances did analysis of the superficial or deep ROI produce better classification performance than analysis of the same MRI parameter in the full-thickness ROI. For example, when samples were grouped by OARSI grade, DW with b = 2000 s/mm² in the superficial ROI resulted in a classification accuracy of 0.67, which is much higher than the poor accuracies of 0.51 and 0.54 observed for the deep and full-thickness ROIs, respectively (see Supplementary Material 2 and 3).

DISCUSSION

There is no recognized direct clinical standard available to characterize the cartilage matrix loss associated with OA.^26,27 The available arthroscopic and radiographic assessments for diagnosing OA are subject to inherent limitations. These limitations include the inability to visualize all locations in the joint, as is possible using 3D MR imaging.⁷ The clinical gold standard of arthroscopic evaluation, such as the Noyes or the Outerbridge classification, requires an invasive procedure, is subjective, and lacks the ability to detect OA prior to chondral softening or blistering given an intact cartilage surface.^28,29 Noninvasive, semi-quantitative radiographic evaluations, such as the Kellgren Lawrence (KL) grade, or MRI assessments, such as the Boston Leeds OA Knee Score (BLOKS), have been shown to correlate well to arthroscopy.³⁰ However, these metrics are intrinsically limited to evaluation of structural changes such as traumatic chondral lesions, osteophyte development, and joint space narrowing, which occur at later stages in OA.³¹ With the goal of early detection to monitor clinical endpoints and to develop therapeutic interventions, quantitative MRI measurements sensitive to cartilage matrix composition should be evaluated as they are complementary to morphological assessments.^11,36–40

Histology is the gold standard for assessing matrix-level changes associated with disease progression and serves as an appropriate ex vivo reference for comparison with MRI.^26,41 Previous work has compared a qualitative scoring based on MRI-derived morphology and SI of osteoarthritic cartilage with Mankin-based histological scoring. For example, work by Gahunia et al. demonstrated encouraging classification accuracies in a nonhuman primate model.⁴² However, the work by Bittersohl et al. reports poor classification accuracy for detecting early histological stages of OA in humans, suggesting a need for more sensitive quantitative measures.⁴³ Building upon earlier methods, the OARSI histopathology assessment system was developed specifically to increase dynamic range over the disease process and improve stratification within early and mild OA states¹⁸, thereby improving detection of matrix-level changes that precede development of structural OA identifiable by morphological imaging.²⁶ Further, the OARSI assessment has demonstrated higher reliability and reproducibility compared to earlier methods.^18,44,45 In order to eliminate any influence of subjectivity involved in MRI assessment as seen in these previous studies, we used quantitative intensity-based MRI measurements for comparison with histological assessment. The fragility of the samples post-harvest necessitated the use of adjacent paired histology and imaging samples to accurately characterize the cartilage tissue.

While the endpoint of OA has been clearly defined clinically, the progression of OA at the early stages as well as any definition of a clinically-relevant histological grade defining early OA is still under investigation¹⁰. The ambiguity of the progression of early OA is due in part to the concurrent degeneration and proliferation prior to appreciable cartilage loss (OARSI grade 4) when OA is suggested to be reversible.^10,26,27 Histological evaluation provides a single point of reference in the dynamic process of cartilage matrix changes even within functionally healthy knees thus limiting its ability to fully characterize OA progression. In this work, normal cartilage was defined based on the range of histological grade and score observed in plugs obtained from the donated body program; these plugs were screened for radiologic and pathologic signs of OA, yet yielded histological scores within the range of matrix damage believed to be reversible, suggesting that these samples are not necessarily representative of progressive OA.²⁶

In devising an MRI-based supervised classification model, we chose to examine both grade and score for several reasons. One argument for using grade is that it directly reflects the severity of matrix-specific changes. However, grade represents the greatest extent of damage to the matrix regardless of the size of the affected area and hence might lead to a lack of correspondence when compared to MRI where measurements are often averaged over larger volumes of tissue both in-plane and through-plane. The score, which includes the amount of affected tissue, might be considered important for such comparisons. An additional consideration in the context of our study is that adjacent samples were examined with histology and MRI, potentially increasing the influence of the stage information contained in the score.

In our study, it was unexpected that T₁W contrast would provide the best classification results for samples grouped by OARSI grade, given the limited variation in cartilage T₁ values observed as a function of pathomemetic degradation.⁴⁶ We have recently shown high classification accuracy when using T₁ to distinguish between normal and enzymatically degraded BNC at both preclinical (9.4 T) and clinical (3 T) field strengths, and under reduced (4°C) and in vivo (37°C) sample temperatures.^15,16 In spite of this high classification accuracy of T₁ observed at 3 T and 37°C, T₁W was one of the poorest classifiers in BNC, with an accuracy of 0.45.¹⁶ These differences in classification results between experimental conditions underscore the importance of the current work which uses human articular cartilage, clinical field strength, and in vivo sample temperature. Although T₁W contrast was one of the best grade-based classifiers in our current study of articular cartilage, its accuracy is low for a clinically viable test.

The best grade-based classification results were obtained with a wide variety of MRI measurements (Table 3) including T₁, T₂, MT, and DW contrasts. In comparison, the best score-based classifiers (Table 4) were 3D spin echo T₂ and several of the 2D spin echo contrasts. The 2D T₂ relaxation time fit, which uses fewer echoes than the fit to the corresponding 3D data, resulted in poorer classification accuracy than that obtained from the 3D acquisition. There are several advantages to using a 3D spin echo acquisition, including higher SNR, the elimination of slice imperfections, and reduction in diffusion effects. Transverse relaxation is influenced by matrix composition, water content, and collagen fiber orientation and integrity.^11,47,48 With advancing OARSI stage, the heterogeneous changes to which T₂ is sensitive, are more pronounced throughout the tissue volume, possibly explaining why measurements of T₂ value as well as intensity in images with heavier T₂ contrast would be better classifiers for score-based, rather than grade-based, analysis.

Overall, the best classification result for all ROI, MRI, and histological metrics studied was the full-thickness T₂W contrast with TE = 60 ms grouped by OARSI score, which yielded a classification accuracy of 0.75. It is quite remarkable that a single contrast-weighted image has the capability of discriminating with such high accuracy and furthermore that it outperforms other quantitative MRI parameters, such as T2 relaxation time, which have previously shown good correspondence to matrix-level changes in cartilage. Given the modest acquisition time needed to obtain a single T₂W image in comparison to that required for obtaining the data needed for measuring quantitative parameters (e.g., relaxation times and ADC), these findings provide compelling support for exploration of images with other contrasts which can be rapidly acquired and for which normalized SI provides high diagnostic accuracy. Direct classification based on SIs in vivo would, however, require establishment of a reference standard for normalization. In addition to external standards applied during scanning, fat, muscle, or synovial fluid within the joint could be used as such a standard.⁴⁹ In light of the high classification accuracy of T₂ and various T₂-weightings in the current study, it would be important to evaluate the influence, if any, of T₂ anisotropy on classification accuracy.^14,50 It is also notable that the analysis using full-thickness ROIs showed the best classification accuracy as this simplifies the assessment of cartilage in vivo by avoiding the need to divide the cartilage thickness into depth-wise zones—which is particularly valuable for real subjects with thin cartilage.

While the classification accuracies obtained in the current work suggest the potential for improved diagnostic value, further work remains to improve the accuracy of quantitative MRI in diagnosing OA. Improvement in classification accuracy could be made by expanding the number of MRI measurements used to build the classification model. This has been previously demonstrated using both multi-parametric classification approaches and multivariate analyses.^51,52 There are many alternate classification methods including fuzzy approaches, which would afford the possibility to assign a probability of disease rather than a binary assignment to a group.⁵² An additional extension of such a multivariate analysis could also incorporate a more complete evaluation of the osteochondral unit including metrics such as cartilage thickness, joint spacing, as well as subchondral and cortical bone structural integrity.⁵³ Further work incorporating combinations of rapid acquisition methods is needed for applying these multivariate approaches, with the aim of achieving a clinically relevant exam time.

In conclusion, we have demonstrated that clinically available MRI measurements have the ability to discriminate between two histologically distinct groups of human cartilage samples using Mahalanobis distance assignment rules. This analysis also demonstrated that high classification accuracy can be obtained using images acquired with a single T₂W image without the need to compute relaxation time maps.